A Model-Based Joint Feature Fusion framework for limited-view photoacoustic tomography reconstruction

In photoacoustic tomography (PAT) reconstruction, imaging systems with limited-view detection geometries capture acoustic signals over restricted angular ranges, which results in reconstruction artifacts and degraded image quality. Model-based iterative methods can effectively compensate for limited-view artifacts, but their iterative optimization procedures remain computationally demanding. While model-based learning methods leverage neural networks to substantially improve reconstruction quality and accelerate convergence, they remain computationally expensive due to repeated evaluations of the physical forward model. Consequently, their applicability in real-time or near-real-time scenarios is limited. To overcome this constraint, this study proposes a novel overall framework: the Model-Based Joint Feature Fusion (MB-JFF) framework, which is developed under the assumption of a homogeneous and lossless acoustic medium. This framework combines the powerful feature extraction capabilities of a joint feature fusion module with the priors embedded in a model-based module. The joint feature fusion module effectively fuses information from both the original sensor data and the model-driven residual sensor data. By integrating these two modules, the proposed method effectively combines the strengths of deep learning with guidance from the forward model, achieving high-quality reconstruction with only a single iteration. Furthermore, we introduce a learned hybrid forward operator that serves as a geometry-informed surrogate forward model within the MB-JFF framework. Through leveraging geometric mapping and a learned wave propagation refiner, this operator rapidly predicts sensor data that closely approximate those obtained from full-wave numerical solvers, while substantially reducing the associated computational cost. Comparative experimental results validate the exceptional performance of the MB-JFF framework. Our proposed approach demonstrates superior reconstruction quality while achieving a notable reduction in computational time compared with other model-based reconstruction methods.